JP5302092B2 - Speech recognition model parameter creation device, speech recognition model parameter creation method, and speech recognition device - Google Patents

Description

  The present invention relates to a speech recognition model parameter creation device, a speech recognition model parameter creation method, and a speech recognition device, and in particular, a speech recognition model parameter creation device that creates speech recognition model parameters using multi-condition learning, and a speech recognition model parameter creation. The present invention relates to a method and a speech recognition apparatus.

  There is multi-condition learning as a conventional speech recognition method. In this multi-condition learning, speech recognition model parameters are learned using speech data having noise in various environments. When the input voice data has noise similar to the environmental noise used in learning, the voice recognition rate is improved.

  A technique for performing speech recognition using such multi-condition learning is disclosed in, for example, Japanese Patent Application Laid-Open No. 2008-122927 (Patent Document 1).

JP 2008-122927 A

  Here, in multi-condition learning, the speech recognition rate decreases if the input speech data is significantly different from the environmental noise used in the learning. That is, multi-condition learning is a speech recognition method that is effective only when it is similar to environmental noise used during learning.

  In addition, even when speech recognition model parameters are learned using speech data having noise in various environments, when performing speech recognition, There is a risk of mixing unknown noise that is very different from noise. In this case, the voice recognition rate decreases.

  An object of the present invention is to provide a speech recognition model parameter creation device capable of improving the speech recognition rate.

  Another object of the present invention is to provide a speech recognition model parameter creation method capable of improving the speech recognition rate.

  Still another object of the present invention is to provide a speech recognition apparatus capable of improving the speech recognition rate.

  The speech recognition model parameter creation device according to the present invention includes a feature amount calculation unit that calculates a feature amount of speech data on which a plurality of noises are superimposed, and a normalization unit that normalizes the feature amount calculated by the feature amount calculation unit And creating means for creating speech recognition model parameters under a plurality of noises using the feature amount normalized by the normalizing means.

  Preferably, the normalization unit includes a filtering unit that filters the feature amount calculated by the feature amount calculation unit using a bandpass filter.

  More preferably, the normalizing means includes dividing means for dividing the feature quantity filtered by the filtering means by the maximum amplitude value.

  More preferably, the creating means creates the speech recognition model parameter by learning.

  In another aspect of the present invention, a method for creating a speech recognition model parameter is used to calculate a feature amount of speech data on which a plurality of noises are superimposed, normalize the calculated feature amount, and normalize the feature amount Is used to create a speech recognition model parameter under a plurality of noises.

  According to still another aspect of the present invention, the speech recognition apparatus includes a recognition unit that performs speech recognition using a speech recognition model parameter created by any of the speech recognition model parameter creation devices described above. .

  The speech recognition model parameter creation device according to the present invention normalizes the feature amount of speech data on which a plurality of noises are superimposed, and creates a speech recognition model parameter using the normalized feature amount. Thereby, since a plurality of noises can be generalized, when performing speech recognition, it can be applied to various noises. As a result, the voice recognition rate can be improved.

  The speech recognition model parameter creation method according to the present invention normalizes the feature amount of speech data on which a plurality of noises are superimposed, and creates a speech recognition model parameter using the normalized feature amount. Thereby, since a plurality of noises can be generalized, when performing speech recognition, it can be applied to various noises. As a result, the voice recognition rate can be improved.

  Further, since the speech recognition apparatus according to the present invention can perform speech recognition using the speech recognition model parameters created by such a speech recognition model parameter creation device, the speech recognition rate can be improved. .

It is a block diagram which shows the structure of a speech recognition apparatus. It is a block diagram which shows the structure of the speech recognition model parameter production apparatus which concerns on one Embodiment of this invention. It is a flowchart shown about the case where a speech recognition model parameter is produced using the speech recognition model parameter creation apparatus, and is memorize | stored in the memory | storage part of a speech recognition apparatus. It is a graph which shows the calculated feature-value. It is a graph shown about the case where the feature-value shown in FIG. 4 is filtered. It is a graph shown about the case where the feature-value shown in FIG. 5 is divided by the maximum amplitude value. It is a flowchart shown about the case where speech recognition is performed using a speech recognition apparatus. It is a flowchart shown about other embodiment in the case of producing | generating a speech recognition model parameter using the speech recognition model parameter creation apparatus, and making it memorize | store in the memory | storage part of a speech recognition apparatus.

  A speech recognition model parameter creation device according to an embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus 10. FIG. 2 is a block diagram showing the configuration of the speech recognition model parameter creation device 20 according to one embodiment of the present invention. First, the configuration of the speech recognition apparatus 10 will be described with reference to FIG.

  The speech recognition device 10 performs speech recognition in the speech recognition device feature amount calculation unit 11 that calculates the feature amount of the input speech data, the recognition unit 12 that performs speech recognition, and the recognition unit 12 via the microphone 14. And a storage unit 13 for storing voice recognition model parameters used at the time.

  The recognition unit 12 performs speech recognition using the feature amount calculated by the speech recognition device feature amount calculation unit 11 and the speech recognition model parameter stored in the storage unit 13. The storage unit 13 stores voice recognition model parameters used when the recognition unit 12 performs voice recognition. The speech recognition model parameter is created by the speech recognition model parameter creation device 20 shown in FIG.

  Next, the configuration of the speech recognition model parameter creation device 20 will be described with reference to FIG. The speech recognition model parameter creation device 20 receives an input of noiseless speech data that does not include noise via the microphone 24, and creates a noise superimposed data by superimposing noise on the received noiseless speech data. 22, a creation device feature amount calculation unit 21 that calculates a feature amount of noise superimposed data created by the noise superimposition unit 22, and a speech recognition model parameter using the feature amount calculated by the creation device feature amount calculation unit 21. And a holding unit 25 for holding a plurality of noise data to be superimposed in the noise superimposing unit 22.

  The noise superimposing unit 22 extracts predetermined noise data from the holding unit 25. Then, noise superimposed data is created by superimposing extracted noise on noiseless voice data.

  The creation device feature value calculator 21 calculates the feature value of the noise superimposed data. As this feature amount, for example, MFCC (Mel Frequency Cepstial Coefficient) can be adopted.

  The learning unit 23 creates a speech recognition model parameter. As the speech recognition model parameter, for example, an average value, a variance value, a transition probability, a weighting coefficient, or the like of a hidden Markov model (HMM) can be adopted.

  The holding unit 25 holds a plurality of noise data. The plurality of noise data includes, for example, various types of environmental noise such as hustle noise, white noise, and factory noise.

  The speech recognition model parameter creation device 20 creates speech recognition model parameters used by the speech recognition device 10. Here, a case where a speech recognition model parameter is created using the speech recognition model parameter creation device 20 and stored in the storage unit 13 of the speech recognition device 10 will be described. FIG. 3 is a flowchart illustrating a case where a speech recognition model parameter is created using the speech recognition model parameter creation device 20 and stored in the storage unit 13 of the speech recognition device 10. This will be described with reference to FIGS.

  First, the speech recognition model parameter creation device 20 receives an input of noiseless speech data via the microphone 24 (step S11 in FIG. 3 and the following steps are omitted). Then, the noise superimposing unit 22 extracts the noise data from the holding unit 25 and superimposes the extracted noise data on the noiseless voice data to create the noise superimposing data. Further, white noise data is extracted, and white noise superimposed data is created in the same manner as described above. Also, factory noise data is extracted and factory noise superimposed data is created in the same manner as described above. In this way, a plurality of noise data are extracted, and a plurality of noise superimposed data are created (S12).

  Then, the creation device feature value calculation unit 21 calculates the feature values of the noise superimposed data created in S12, that is, the hustle noise superimposed data, the white noise superimposed data, and the factory noise superimposed data (S13). Specifically, in the noise superimposition data, 20 to 30 ms is set as one frame, and the noise superimposition data is divided into a plurality of frames, thereby calculating a feature amount in each frame. In the division, the division is performed so that the one frame and the subsequent frame partially share data. Here, the creation device feature value calculation unit 21 operates as a feature value calculation unit.

  FIG. 4 is a graph showing the calculated feature amount. With reference to FIG. 4, the dotted line indicates the feature amount of the hustle noise superimposed data, the alternate long and short dash line indicates the feature amount of the white noise superimposed data, and the solid line indicates the feature amount of the factory noise superimposed data. The horizontal axis indicates the frame, and the vertical axis indicates the amplitude value of the feature amount. Referring to FIG. 4, in the non-speech section A, the difference in amplitude between the feature amount of the white noise superimposed data, the feature amount of the hustle noise superimposed data, and the feature amount of the factory noise superimposed data is large. Also in the speech section B, the difference in amplitude is large between the feature amount of the white noise superimposed data, the feature amount of the hustle noise superimposed data, and the feature amount of the factory noise superimposed data. Further, when comparing the amplitude values of the feature values in the frame 10, the noise superimposition data indicates 0.8, the white noise superimposition data indicates -4, and the factory noise superimposition data indicates -0.3. Yes.

  Then, when the calculation of the feature amount is completed, the calculated feature amount is normalized to obtain a normalized feature amount, that is, a normalized feature amount. This normalized feature value utilizes the fact that the time change of the voice is gradual. Specifically, first, a feature amount is filtered using a bandpass filter (S14). That is, the feature amount is filtered so that only frequency components in a predetermined range are allowed to pass and other frequency components are not allowed to pass. The bandpass filter is a FIR (Finite impulse response) type filter. Thereby, a process can be performed stably.

  FIG. 5 is a graph showing the case where the feature amount shown in FIG. 4 is filtered. As in FIG. 4, the dotted line indicates the feature amount of the hustle noise superimposed data, the dashed line indicates the feature amount of the white noise superimposed data, and the solid line indicates the feature amount of the factory noise superimposed data. The horizontal axis indicates the frame, and the vertical axis indicates the amplitude value of the feature amount. Referring to FIG. 5, in the non-voice section A and the voice section B, the feature amount of the white noise superimposed data, the feature amount of the hustle noise superimposed data, and the feature amount of the factory noise superimposed data are as shown in FIG. However, the waveforms are uniform and the difference in amplitude is small. Further, when comparing the amplitude values of the feature values in the frame 10, −0.3 is shown for the hustle noise superimposed data and the factory noise superimposed data, and −0.7 is shown for the white noise superimposed data. That is, by filtering, it is possible to align waveforms in a plurality of data having different waveforms, and generalize the data.

  Then, when filtering of the feature amount is completed, the filtered feature amount is divided (divided) by the maximum amplitude value of the filtered feature amount (S15). For example, referring to FIG. 5, in the hustle noise superimposition data, the maximum amplitude value is divided by a, and in the white noise superimposition data, the maximum amplitude value is divided by b. Is divided by c which is its maximum amplitude value.

  FIG. 6 is a graph showing the case where the feature amount shown in FIG. 5 is divided by the maximum amplitude values a, b, and c. Similar to FIG. 5, the dotted line indicates the feature amount of the hustle noise superimposed data, the alternate long and short dash line indicates the feature amount of the white noise superimposed data, and the solid line indicates the feature amount of the factory noise superimposed data. The horizontal axis indicates the frame, and the vertical axis indicates the amplitude value of the feature amount. Referring to FIG. 6, in the non-voice section A and the voice section B, the feature amount of the white noise superimposed data, the feature amount of the hustle noise superimposed data, and the feature amount of the factory noise superimposed data are as shown in FIG. However, the waveforms are more uniform and the difference in amplitude is smaller. In addition, when comparing the amplitude values of the feature values in the frame 10, it shows -0.23 in the hustle noise superimposition data and factory noise superimposition data and -0.27 in the white noise superimposition data. That is, by dividing by the maximum amplitude value, it is possible to further align the waveforms in a plurality of data having different waveforms, and generalize the data.

  In this way, by filtering using a bandpass filter and dividing by the maximum amplitude value, the calculated feature value is normalized to obtain a normalized feature value. Here, the creation device feature amount calculation unit 21 operates as a normalization unit, a filtering unit, and a division unit.

  Then, the learning unit 23 creates a speech recognition model parameter using each normalized feature amount (S16). Specifically, a speech recognition model parameter is created by performing multi-condition learning. Here, the learning unit 23 operates as a creation unit. Then, the speech recognition model parameters are stored in the storage unit 13 of the speech recognition device 10 (S17).

  As described above, the speech recognition model parameter creation device 20 normalizes the feature amount of speech data on which a plurality of noises are superimposed, and creates a speech recognition model parameter using the normalized feature amount. Thereby, since a plurality of noises can be generalized, when performing speech recognition, it can be applied to various noises. As a result, the voice recognition rate can be improved.

  In addition, such a speech recognition model parameter creation method normalizes the feature amount of speech data on which a plurality of noises are superimposed, and creates a speech recognition model parameter using the normalized feature amount. Thereby, since a plurality of noises can be generalized, when performing speech recognition, it can be applied to various noises. As a result, the voice recognition rate can be improved.

  Note that the amplitude values of the feature amounts shown in FIGS. 4 to 6 vary depending on the input audio data.

  Next, a case where voice recognition is performed using the voice recognition device 10 will be described. FIG. 7 is a flowchart illustrating a case where voice recognition is performed using the voice recognition device 10. This will be described with reference to FIGS.

  First, the speech recognition apparatus 10 receives input of speech data via the microphone 14 (S21). Then, the voice recognition device feature quantity calculation unit 11 calculates the feature quantity of the voice data.

  The calculation of the feature amount is performed in the same manner as S13 to S15 in FIG. That is, the audio data is divided into a plurality of frames, and the feature amount is calculated in each frame (S22). When the calculation of the feature value of the voice data is completed, the calculated feature value is normalized, that is, filtered using a bandpass filter, and divided by the maximum amplitude value, thereby obtaining the normalized feature value of the voice data. (S23).

  Then, the recognition unit 12 performs voice recognition using the normalized feature amount of the voice data calculated in S23 and the voice recognition model parameters stored in FIG. 3 described above (S24). Here, the recognition unit 12 operates as a recognition unit. For example, the speech recognition is performed based on the calculated likelihood value by comparing the normalized feature amount of the speech data calculated in S23 with the speech recognition model parameter.

  As described above, since the speech recognition apparatus 10 can perform speech recognition using the speech recognition model parameters created by the speech recognition model parameter creation apparatus 20, the speech recognition rate can be improved.

  In the above-described embodiment, when the speech recognition model parameter is created using the speech recognition model parameter creation device 20 and stored in the storage unit 13 of the speech recognition device 10, as shown in S14 to S15. Although an example of filtering using a bandpass filter and dividing by the maximum amplitude value has been described, the present invention is not limited to this, and after filtering using a bandpass filter, division by a variance value may be performed.

  In the above embodiment, when the speech recognition model parameters are created using the speech recognition model parameter creation device 20 and stored in the storage unit 13 of the speech recognition device 10, as shown in S14 to S15. Although an example of filtering using a bandpass filter and dividing by the maximum amplitude value has been described, the present invention is not limited to this, and other embodiments described below may be employed.

  FIG. 8 is a flowchart illustrating another embodiment in which a speech recognition model parameter is created using the speech recognition model parameter creation device 20 and stored in the storage unit 13 of the speech recognition device 10. Note that S31 to S33 are the same as S11 to S13 shown in FIG.

  Referring to FIG. 8, first, when the calculation of the feature value is completed in S33, an average value of the calculated feature values is obtained (S34). Next, the obtained average value is subtracted from the feature amount calculated in S33 (S35). Then, the subtracted feature amount is filtered using a low-pass filter (S36). Further, the filtered feature value is divided by the maximum amplitude value of the filtered feature value (S37).

  In this way, the average value is subtracted, filtered using a low-pass filter, and divided by the maximum amplitude value, thereby normalizing the calculated feature value to obtain a normalized feature value. Then, a speech recognition model parameter is created (S38) and stored in the storage unit 13 (S39).

  In the above embodiment, when the speech recognition model parameters are created using the speech recognition model parameter creation device 20 and stored in the storage unit 13 of the speech recognition device 10, as shown in S14 to S15. Although an example of filtering using a band pass filter and dividing by the maximum amplitude value has been described, the present invention is not limited to this, and it may be only filtered using a band pass filter.

  In the above-described embodiment, an example in which an FIR type filter is used has been described. However, the present invention is not limited to this, and an IIR (Infinite Impulse Response) type filter may be used. Thereby, the processing amount of processing can be reduced.

  In the above-described embodiment, an example has been described in which the crowd noise superimposition data, the white noise superimposition data, and the factory noise superimposition data are generated in S12 to S13, and the respective feature amounts are calculated. In addition, the configuration may be such that factory noise superimposition data is not created. That is, the feature amount of at least one of the hustle noise superimposed data and the white noise superimposed data may be calculated.

  In the above-described embodiment, the example of creating the hustle noise superimposing data, the white noise superimposing data, and the factory noise superimposing data as the noise superimposing data in S12 has been described. It may be configured to include data in which the amount of noise is made small enough to be ignored. In other words, noiseless data may be included as noise superimposed data.

  In the above-described embodiment, the example in which the holding unit 25 holds a plurality of noise data in various environments has been described. However, the present invention is not limited to this example. You may keep the noise. That is, the plurality of noises are configured to include a plurality of specific types of noises. For example, a plurality of noises related to a factory as a specific type, specifically, a first factory noise and a second factory noise are held. And the normalization feature-value in a factory is obtained by producing the noise superimposition data of a 1st factory, and the noise superimposition data of a 2nd factory. In addition, a plurality of noises related to the hustle and bustle, specifically, the first hustle noise and the second hustle noise are retained as the specific type, and the first hustle noise superimposition data and the second hustle noise are stored. By creating the noise superimposition data, the normalized feature amount in the crowd is obtained. Then, the speech recognition model parameter may be created using the normalized feature value in the factory and the normalized feature value in the hustle and bustle.

  In the above embodiment, the holding unit 25 has been described as an example including noise data such as hustle noise, white noise, and factory noise. It may be configured to include noise data such as a plurality of room noises and audio sounds, and can be arbitrarily set.

  Further, the speech recognition model parameter creation device 20 may be implemented by hardware or software. Similarly, the speech recognition apparatus 10 may be implemented by hardware or software.

  As mentioned above, although embodiment of this invention was described with reference to drawings, this invention is not limited to the thing of embodiment shown in figure. Various modifications and variations can be made to the illustrated embodiment within the same range or equivalent range as the present invention.

  DESCRIPTION OF SYMBOLS 10 speech recognition apparatus, 11 speech recognition apparatus feature-value calculation part, 12 recognition part, 13 memory | storage part, 14 microphone, 20 speech recognition model parameter creation apparatus, 21 creation apparatus feature-value calculation part, 22 noise superimposition part, 23 learning part, 24 microphone, 25 holding part.

Claims (7)

Pre retained, and a plurality of noise data, the feature amount calculating means for calculating a plurality of audio data each feature amount created by Rukoto superimposed on noise-free data corresponding to a plurality of noise,
Normalizing means for normalizing each feature quantity calculated by the feature quantity calculating means in order to generalize the plurality of noises,
The normalization means includes filtering means for filtering each feature quantity calculated by the feature quantity calculation means using a bandpass filter so as to pass only frequency components in a predetermined range,
A speech recognition model parameter creation device further comprising creation means for creating a speech recognition model parameter under the plurality of noises using each feature amount normalized by the normalization means. Pre retained, and a plurality of noise data, the feature amount calculating means for calculating a plurality of audio data each feature amount created by Rukoto superimposed on noise-free data corresponding to a plurality of noise,
Normalizing means for normalizing each feature quantity calculated by the feature quantity calculating means in order to generalize the plurality of noises,
The normalizing means includes
Subtracting means for subtracting an average value of the feature values from each feature value calculated by the feature value calculating means;
Means for filtering the feature amount after subtraction by the subtracting means using a low-pass filter,
A speech recognition model parameter creation device further comprising creation means for creating a speech recognition model parameter under the plurality of noises using each feature amount normalized by the normalization means.   The speech recognition model parameter creation device according to claim 1, wherein the normalization unit further includes a division unit that divides the feature amount filtered by the filtering unit by a maximum amplitude value or a variance value thereof.   The speech recognition model parameter creation device according to claim 1, wherein the creation unit creates the speech recognition model parameter by learning. A method executed by a speech recognition model parameter creation device,
Pre retained the steps plurality of noise data corresponding to a plurality of noise, which calculates a plurality of audio data each feature amount created by Rukoto superimposed on noiseless data,
In order to generalize the plurality of noises, a normalization step of normalizing each calculated feature amount, and
The normalizing step includes a step of filtering each calculated feature amount using a band-pass filter so as to pass only frequency components in a predetermined range.
A speech recognition model parameter creation method, further comprising creating speech recognition model parameters under the plurality of noises using each normalized feature amount. A method executed by a speech recognition model parameter creation device,
Pre retained the steps plurality of noise data corresponding to a plurality of noise, which calculates a plurality of audio data each feature amount created by Rukoto superimposed on noiseless data,
In order to generalize the plurality of noises, a normalization step of normalizing each calculated feature amount, and
The normalizing step includes
Subtracting an average value of the feature quantities from the calculated feature quantities;
Filtering the feature amount after subtraction using a low-pass filter,
A speech recognition model parameter creation method, further comprising creating speech recognition model parameters under the plurality of noises using each normalized feature amount.   A speech recognition apparatus comprising a recognition means for performing speech recognition using the speech recognition model parameter created by the speech recognition model parameter creation apparatus according to claim 1.

Images